Skip to main content
Have a project in mind? Let's talk
Book a call Tell us about your project Book a Call
AGENT
  AI Agents

AI Agents and Claude Automation. Built With a Person in the Loop.

We build and run AI agents on Claude that read your messages and documents, use your tools and finish multi-step jobs, with approval points where a person decides and logs you can read. We use Claude Code where it fits to build internal tools and integrations quickly, and write the CLAUDE.md files that make agents follow your company's rules.

ClaudeOur main model for agents
ApprovalPoints by design
24hReply time
AI Agents Claude Automation Claude Code Builds CLAUDE.md Files MCP Connections Approval Points Run Logs Make and n8n AI Agents Claude Automation Claude Code Builds CLAUDE.md Files MCP Connections Approval Points Run Logs Make and n8n

What an AI Agent Does That a Script Does Not

A script follows the steps you gave it. An agent works out the steps from the situation, so it needs tighter limits, not looser ones.

Devvista is a software development and growth agency. We build AI agents on Claude for businesses whose work arrives as messy text: an enquiry that needs qualifying, a supplier invoice in an odd layout, a support message that could belong to three teams. Anthropic's engineering guide separates two kinds of system. Workflows are orchestrated through predefined code paths. In agents, the model dynamically directs its own process and tool use. Both are useful, and the guide advises starting with the simplest solution and adding complexity only when it pays off.

That advice is how we scope. An agent costs more to run and to test than a fixed workflow, and it can compound its own mistakes across steps, so we use one only when the number of steps cannot be written down in advance. Everything else gets a rule-based flow, a single model call, or a script.

Where the three approaches fit. Use the table to see which one your problem needs before you read further.

ApproachHow it decidesBest forMain risk
Rule-based workflowFixed steps and conditions you definePredictable work between systems with APIsBreaks on input nobody planned for
Single model step inside a workflowRules decide the path, the model reads or classifies one inputSorting, extracting and summarising textA wrong label passes silently without a review rule
AI agentThe model chooses the next step and the tool to useOpen-ended jobs where the steps depend on what it findsCost, latency and compounding errors without limits

Agents and Tools We Build for Businesses

Each build starts with one job written down step by step, and one person who owns the result.

01

Lead intake and qualification agents

An agent reads a new enquiry, checks your CRM for an existing contact, scores it against your qualification rules, drafts a reply and hands the lead to the right person. A team member approves anything that is sent to the prospect.

Leads
02

Support triage

Incoming messages are read, tagged, routed to the right queue and given a draft answer from your own help content, with unclear cases sent to a person.

Triage
03

Document and invoice handling

Fields pulled from invoices, contracts and forms, checked against your records and filed, with mismatches queued for review.

Docs
04

Reporting and data clean-up

Weekly summaries written from the numbers in your systems, and duplicate or inconsistent records found, explained and fixed in batches you approve before anything is overwritten.

Reports
05

Internal tools built with Claude Code

Admin panels, import scripts, dashboards and small integrations built, where it fits, with Claude Code, then reviewed, tested and handed over as ordinary code your own developers can read.

Tools
06

MCP integrations

Connections that let an agent read from and act in your CRM, email, accounting tool or database through one standard interface, with permissions set per tool. Where no connector exists, we build one on top of your APIs.

MCP

Claude Code, CLAUDE.md, MCP and Skills in Plain Terms

These are the parts we work with. Each one has a job, and none of them is magic.

Anthropic describes Claude Code as an agentic coding tool that reads your codebase, edits files, runs commands and integrates with your development tools. We use it as a fast pair of hands on internal tools, scripts and integrations. It does not replace review: a developer of ours reads and tests what it produces before it is handed to you.

The table lists the pieces that matter for a business project, with what Anthropic's documentation says each one is and what it means in your build.

PieceWhat the documentation saysWhat it means for your project
CLAUDE.mdA markdown file in the project root that Claude Code reads at the start of every session, for standards, decisions and checklistsYour company's rules are written once and applied every run, and you can read and edit them
MCPAn open standard for connecting AI applications to external data sources, tools and workflowsOne consistent way for an agent to reach your CRM, mail, accounting and database
SkillsPackaged repeatable workflows a team can share, such as a review or deploy routineA routine your team repeats becomes a named, reusable step
HooksShell commands that run before or after Claude Code actionsChecks such as formatting or linting run automatically around the agent's work
SubagentsMultiple agents working on parts of a task, with a lead agent coordinating and merging resultsA larger job can be split, but only when the extra moving parts are worth it
Agent SDKA toolkit for building custom agents with control over orchestration, tool access and permissionsThe route for an agent that runs inside your own product or back office

Claude Code can also be scripted from the command line and run on a schedule, which is how recurring checks and reports get built. The Claude Code for business automation guide shows what that looks like day to day, and the CLAUDE.md file explained post walks through what goes in the file. For the connection layer, read how MCP connects Claude to business tools.

From First Agent
to Handover

You see the agent working on test data every week before it touches live records.

01
Step One

Scope one first agent

We pick a single job that repeats often, with a clear definition of a good result. After a short call we send a proposal for that first piece, not for a programme.

02
Step Two

Set up accounts in your name

Model, hosting and tool accounts are created under your company, and the agent gets its own narrow credentials. We sign an NDA first if you want one.

03
Step Three

Write the rules down

Your policies, tone and limits go into a CLAUDE.md file and project instructions. You review them in plain language, so the rules are yours, not ours.

04
Step Four

Build on staging, demo weekly

The agent runs on a staging copy with test data. Each week you watch it work, and we change the rules and tools from what you see.

05
Step Five

Add approval points and logs

Sending, paying, deleting and editing customer records wait for a person. Every run is logged so you can see what the agent read, which tool it used and what it decided.

06
Step Six

Hand over or keep it running

You receive the code, instructions, accounts and a short runbook. We can keep running and improving the agent, or your team takes over, whichever you prefer.

How We Keep an Agent Inside Its Limits

The useful question is not whether the agent will ever be wrong. It is what the worst wrong action could be, and whether it is even possible.

Anthropic's guide recommends extensive testing in sandboxed environments along with appropriate guardrails, and suggests agents pause for human feedback at checkpoints or when they hit a blocker. It also suggests stopping conditions such as a maximum number of iterations. We build those ideas into every agent as concrete settings you can read.

GuardrailWhat we doWhy it matters
Least privilegeEach agent gets only the tools and records its job needs, with read-only access firstAn agent that cannot delete a record cannot delete one by mistake
Separate credentialsA dedicated login for the agent, never a staff member's accountActions are attributable and the access can be revoked in one step
Approval pointsAnything sent, paid, deleted or hard to undo waits for a named personThe decision that carries risk stays with someone accountable
Limits on a runA cap on steps, time and spend, then the run stops and reportsA confused agent cannot loop for hours or run up a bill
Review of outputSample checks of finished work, and a human read of anything customer-facing at firstQuality problems show up while the volume is still small
Error handlingFailed steps are retried by rule, then queued with an alert to a named personNothing fails quietly
Staging firstNew rules and tools are tested on copies of your data before live useA bad change is found before a customer sees it

The same guide puts care into tool design as well: clear tool descriptions, defined edge cases and tools built so mistakes are harder to make. We apply that to the tools we give an agent, for example by making a tool accept only the record types it should touch. For process-level reliability such as retries and alerts, the workflow automation page covers the platform side.

When an Agent Is the Wrong Answer

We tell you when something simpler will do. These are the usual cases.

01
The steps never change

If you can write the whole process as fixed rules, a workflow is cheaper, faster and easier to test. Add a model only for the one step that reads text.

02
One wrong action is expensive

Payments, legal commitments and anything hard to reverse should not be left to an agent. It can prepare the case, and a person decides.

03
Your data is not in a usable state

An agent working from duplicate, outdated records produces confident wrong answers. Clean-up comes first, and it can be the first project.

04
The volume is tiny

Ten tasks a month rarely repay the build and the testing. A checklist or a saved template does the job.

05
Nobody will own it

Someone has to read the alerts, approve the actions and decide when the rules change. Without that person, an agent drifts.

How This Differs From Our AI and Workflow Pages

Three pages, three questions: what should be automated, how the steps run, and who decides the steps.
PageStart here ifWhat it covers
AI and automationYou want to find and automate repetitive processes with models and rulesProcess audit, document processing, chatbots and predictive work
Workflow automationThe steps are known and need to run reliably between systemsTriggers, approvals, syncs, alerts and run logs on Make, GoHighLevel or code
AI agents (this page)The work depends on reading context and choosing the next stepClaude agents with tools, Claude Code builds, CLAUDE.md rules and MCP connections

If you are unsure which applies, a short call is enough to sort it out. Not every project needs an agent, and some need a custom build instead of a platform, which is what our custom software development service is for. Weighing a build against doing it in-house is covered in DIY or hire an AI automation team.

For a wider look at the options, read what AI agents can automate for a business and how to automate your business with Claude.

Related Work We Have Already Delivered

These case studies are the closest published work. Some clients are confidential, so those pages describe the system and not the company.

Claude inside an approved workflow. In our Make lead intake automation, a single Make scenario maps a 112-field Typeform submission into GoHighLevel and Airtable, generates a business document with Claude AI, and holds delivery until a team member approves it in Slack. That is the pattern we use for agents: a model does the drafting, and a person holds the release.

A multi-agent system with manual review. The credit repair automation platform runs a 17-agent AI analysis system that flags high-risk cases for manual review.

Rules that gate a sales process. The GoHighLevel lead qualification build is rule-based, not agent-based. We list it because most agent projects begin with the same qualification rules.

We do not publish results figures for these builds beyond what each case study states, and none is a promise about what yours will do. You can browse all of it on the work page.

Platforms &
Tools

We choose per project. Tell us your stack and we will work with it.

Claude Claude Code CLAUDE.md MCP Make n8n GoHighLevel Airtable Slack Webhooks REST APIs Node.js Python PostgreSQL

Where the Tool Facts Come From

Statements on this page about Claude Code, CLAUDE.md, MCP, skills, hooks, subagents and the Agent SDK come from the documentation below, opened in November 2026. Tools change quickly, so check the linked page before you rely on a detail.

Straight
Answers.

Plain answers to what we are asked before an agent is built.

Still have a question?Send it over and we will answer personally.Send a Message

An AI agent is a model that works toward a goal in several steps and decides for itself which tool to use next, such as reading an email, looking up a record or drafting a reply. Anthropic's engineering guide describes agents as systems where the model dynamically directs its own process and tool use, as opposed to workflows that follow predefined code paths. In business use, we pair the agent with tools you approve and checkpoints where a person decides.

A workflow follows steps you wrote down. An agent chooses its steps from the situation in front of it. AI automation is the wider term for putting a model inside a process, for example to read documents or classify messages. Most real projects mix them, and we only use an agent where the number of steps cannot be fixed in advance.

Claude Code is Anthropic's agentic coding tool. Per its documentation it reads a codebase, edits files, runs commands and connects to development tools. We use it to build internal tools, scripts and integrations faster, and every change still goes through human review before it touches your systems.

CLAUDE.md is a markdown file in a project that Claude Code reads at the start of every session, according to Anthropic's documentation. It holds standards, decisions and checklists. For a business it is where your rules live: which systems an agent may touch, how to word customer replies, what always needs approval. Our blog post on the CLAUDE.md file explains the format in detail.

MCP, the Model Context Protocol, is an open standard for connecting AI applications to external data sources, tools and workflows. You need it when an agent has to read from or act in your CRM, email, accounting tool or database through one consistent connection. If the job only needs a script and a single API, plain code is simpler and we will say so.

Only if you decide it should. We start with read access and drafts, and put an approval point before anything is sent, paid, deleted or changed in a customer record. Autonomy is widened one action at a time, after the logs show the agent handling it correctly.

Yours. We set up model, hosting and tool accounts in your company's name, so billing, data and access stay with you. Agents get their own narrow credentials instead of a staff member's login, and you can revoke them at any time.

Yes. When the job is mostly connecting systems with clear rules, a platform such as Make or n8n is often the faster and cheaper build, with a Claude step added only where text or judgment is involved. We recommend per process and tell you when no agent is needed.

Every project is quoted after a short call, because scope, the number of systems and the approval rules change the work a lot. The contact form asks for your budget range so the first reply can be realistic, and the quote is written before any build starts.

We sign a mutual NDA on request before detailed conversations. Code, prompts, CLAUDE.md files and configuration are handed over to you at the end of the project, and the accounts they run on are already in your name.
Let's talk

Tell Us About Your First Agent.

Describe the job you want off your team's desk. We will say whether an agent, a workflow or a simple script fits, and quote it after a short call.

  • 30-minute consultation
  • NDA signed on request
  • Reply within 24 hours
  1. Book a call30 minutes, no pitch. Tell us what you want to build.
  2. Get a straight answerHonest feedback on fit, scope and timeline.
  3. Decide with no pressureMove ahead only if it feels right. No commitment required.
Chat on WhatsApp